Chapter 26 spiritual emergence

spiritual emergence Holland had thought he could finish a book in a year or two, a compilation of graph theorems, genetic algorithms, and his general thinking on fitness.But in fact, the publication of this book took him ten years.His writing and research have always been in tandem, with new ideas to explore or new aspects of theory to analyze.He assigned several graduate students under his supervision to do computer experiments, that is, to verify whether genetic algorithms are really useful and whether they are an effective way to solve optimization problems.Holland felt that he was proposing and practicing his theory of adaptation at the same time, and he wanted to be precise, thorough and precise.

He certainly did. The 1975 book "Adaptation in Natural and Artificial Systems" lays out a lot of equations and analysis.The book summarizes Holland's two decades of thinking on the profound inner connection between learning, evolution and creativity, and makes a thorough statement on genetic algorithms. Holland's work did not resonate in the wider computer science world outside of UM.For those who like elegant, concise, and proven correct algorithms, Holland's genetic algorithm is too eccentric.The artificial intelligence community responded enthusiastically to Holland's genetic algorithm research.In this circle, his book sales can reach one hundred to two hundred copies a year.But even so, and even the occasional review of his books, it's mostly "John is a very smart guy, but..." sort of thing.

Of course, Holland did not force his point of view.He only published a few papers, relatively few papers, and only gave lectures when he was invited, that's all.He did not hype genetic algorithms at major academic conferences, did not apply genetic algorithms to popular application fields such as medical diagnosis that can win research funds and attract attention, and did not strive for huge investment to create genetic algorithm "labs" , nor published a best-selling book calling for the imminent use of genetic algorithms to deploy federal funds to counter the Japanese threat.

In short, he is not at all playing the game of self-promotion in the academic world, which does not seem to be his favorite game.Rather, he doesn't really mind if he wins or loses.Metaphorically speaking, he still prefers to hang out with his friends in the basement and play games."It's like playing baseball, except you're on the non-team instead of the main team," Holland said. "But it's the fun of baseball that counts, not the team. The science I do is important to me. Words are always fun." "I find it annoying if no one wants to listen to me. But I've always been very lucky to have graduate students who are bright and interested in my research topics and I think I can relate to them intellectually."

This really reflects his attitude towards playing games in the basement with his teenage buddies: At Michigan, Holland put a lot of energy into working with the buddies in his immediate circle.Especially when he has six or seven graduate students under his command at any one time, which is far more than the usual number of graduate students supervised by supervisors.In fact, since the mid-1960s, more than one of the graduate students he supervised has received a doctorate every year. "Some of them are really, really smart, and I've had a lot of fun with them," he said.Holland has seen too many professors accumulating long lists of papers that are actually the result of joint research between them and the graduate students they supervise, or even written entirely by their students.So he deliberately mentors graduate students in a rather hands-off manner. “They all do research on their own terms, doing things that interest them. And then once a week we all gather around the table and one of them tells everyone how far he’s gone with his dissertation, and we all There’s critique and discussion about it. It’s a lot of fun for everyone involved.”

In the mid-1970s, Holland and a group of like-minded colleagues in the Faculty started weekly freelance discussions on any question of evolution and adaptation.In addition to Burks, the group included the political scientist Robert Axelrod.Robert tries to understand why and when people cooperate with each other instead of stabbing each other in the back.There were also political scientists Michael Cohen and William Hamilton, who specialized in the social dynamics of human organization.William is an evolutionary biologist who, with Axelrod, studies symbiosis, social behavior and other forms of biological cooperation.

"Mike Cohen was a go-between," Holland recalls.Not long after his book on adaptation was published, Cohen dropped in on one of his classes.He came up to Holland one day after class and introduced himself, saying, "You should really go and talk to Axelrod." Holland did.Through Axelrod, he met Hamilton again.Soon, the members of the BACH team, Boxer, Axelrod, Cohen, and Hamilton joined hands. (BACH is made up of the initials of the four of them. They work together almost always. At the beginning of the group, they tried to include Stuart Kaufman, but Kaufman went to Penn.) "Put What connected us was that we both had strong mathematical backgrounds and a strong sense that the problems of evolution and adaptation were broader than any one single problem. We started meeting regularly: someone would read an article, and Everyone discussed together, which stimulated a lot of exploratory thinking." Holland said.

This is especially true for Holland.He had completed the book on adaptation, but his discussions with members of the BACH group touched on areas left untouched in the book and left to study.He firmly believed that the genetic algorithm and the graph theorem firmly grasped the essential problem of evolution.But even so, he still can't help but regret that the genetic algorithm's naked explanation of evolution is too simple after all.His theory treats "organisms" directly as simple DNA designed by programmers, and there must be shortcomings in such a theory.What does it tell us about the evolution of complex organisms in complex environments?Nothing can be explained.Genetic algorithms are pretty good, but only genetic algorithms themselves, not including adaptive actors.

In this sense, the genetic algorithm is not an imitation of human mental fitness.Because it's computationally too biological to see how complex concepts arise, evolve, and recombine in the human mind.For Holland, this fact was increasingly frustrating.It had been twenty-five years since he first heard Herb's concept, and yet he remained convinced that spiritual adaptation and adaptation in nature were but two different sides of the same coin.Moreover, he still believed that if they were really the same thing, they could be described by one and the same theory. Therefore, from the late 1970s, Holland began to study the theory.

Fundamentally, an adaptive agent is often playing games with its environment.And what exactly does that mean?If stripped down to the essence, what will happen to the survival and development of the players in the game? Holland thinks two things will happen: prediction and feedback.This was an insight he had when he talked to Samuel about checkers while he was working at IBM.Prediction is exactly what the word means: think in advance.Holland still remembers Samuel repeating this over and over again. "The essence of playing a good game of checkers or chess is to bet on less obvious layouts." That is, to play chess that can put you in a favorable position later.Prediction can help you seize opportunities or avoid pitfalls.An agent who can think in advance clearly has an advantage over an agent who cannot think in advance.

But the concept of forecasting is almost as subtle as the concept of building bricks.For example, we often think of prediction as conscious human thinking based on rough simulations of the world.Of course there are many such predictions made through simulations.A supercomputer's simulation of the climate is one example, a company starting a business plan is another, the Federal Reserve's economic planning is another, and even England's Stonehenge is a model for simulation: the surrounding design of the monolith can Let the divining priest predict the coming of the spring and autumn equinoxes like a crude but effective computer.Moreover, models of various simulations are often in our heads.A shopgoer tries to imagine a new sofa in his living room, or a timid employee tries to imagine the consequences of offending his boss.We use these "internal models" all the time.In fact, many psychologists believe that conscious thinking is based on "internal models" of thinking. But to Holland, the concept of predictive and simulation models is actually much deeper than conscious thought.From this point of view, it is far more profound than the existence of the brain.“All complex adaptive systems—economies, minds, organisms, etc.—build models that allow them to predict the world,” he declared. Even bacteria.Many bacteria have a special system of inducible enzymes that make them swim in the direction of stronger glucose concentrations.Undoubtedly, these induced enzymes mimic a key aspect of the bacterial world: chemicals always diffuse outward from a source, in ever-decreasing concentrations as they get farther from the source.Induced enzymes naturally coded in the genetic code a clear prediction: If you swim toward higher concentrations, you're likely to find something nutritious. "It's not an intentional pattern," Holland said. "But organisms that follow this pattern have an advantage over organisms that don't." The same goes for the American viceroy butterfly, Holland said.This butterfly is a striking orange and black insect.Birds would no doubt salivate at it if they had tasted it.But birds rarely prey on these butterflies because their wing pattern has evolved to resemble the unpalatable monarch butterfly that all young birds avoid.So this happened: the DNA of the butterfly butterfly encoded a model that simulated a world where there were birds, there were king butterflies, and the king butterflies tasted bad.Every day, the colorful butterfly flies around among the flowers, no doubt betting its own life on its assumption that its simulation of the outside world is correct. You'll also find the same truth in a variety of different organisms.In the case of a company, Holland said, imagine a factory taking a regular order for, say, 10,000 gadgets.Since it was a regular order, the factory staff probably didn't think much of it.They just follow "regular operating procedures", a series of formal procedures, to produce. "If the situation is ABC, then take action XYZ." Like bacteria and variegated butterflies.Coded into these procedures is the plant's modeled world and predictions about it: "If, in ABC, action XYZ would have a good effect." The existence of such a simulated model is not known.After all, routine operating procedures are often learned by copying, and there are not so many whys to ask.If the plant had been in business for years, no one would probably remember why things had to be done the way they were.But anyway, because normal operating procedures are developed and implemented collectively, the entire plant behaves as if it fully understands the simulation model. Holland said that in the field of cognition, anything we call "technology" or "expert knowledge" is an embedded model, or more precisely, a set of long-term experience accumulated and condensed , the vast interlocking routine of operations engraved on the nervous system.Let an experienced physics teacher look at the textbook exercises, and he will not waste his time copying all the formulas within sight like a novice.The program in his head will always tell him the solution to the problem immediately: "Aha, this is an energy problem." Serve the ball to the tennis star Evert, and she won't waste any time arguing about how to catch the ball. , the programming in her head will immediately let her instinctively catch the ball back and hit you hard. Holland likes to cite the technique of medieval architects who created Gothic cathedrals when talking about the expertise involved.They couldn't calculate strength and bearing capacity, or anything that a modern architect can calculate.The twelfth century had no contemporary physics and structural analysis.Those medieval architects who built those high vaulted ceilings and massive buttresses relied on standard operating procedures passed down from master to disciple, practical experience that gave them common sense about which structures would hold up and which would collapse.In them, the model of physics is entirely implicit and intuitive.Yet the structures invented by these medieval craftsmen are still standing more than a thousand years later. The examples are endless, Holland said. DNA itself is an embedded model, Gene says: "Under such conditions we expect our deliberately chosen organisms to have a chance of developing well." Human culture is an embedded model, a rich and complex mythology Stories and symbols imply people's beliefs about their world and the correctness of their rules of conduct.In this regard, Samuel's computer checkers also contains internal patterns, as it becomes more and more familiar with the opponent's style of chess, it will constantly change the expectations it assigns to various chess choices, thereby forming its own Inner patterns. Indeed, patterns and forecasts are everywhere.But where do patterns come from?How do all systems, natural or artificial, come to know enough about the universe to make predictions about future events?Talking about "consciousness" alone doesn't help, he said.Most models are clearly not conscious: the nutrient-seeking bacteria, for example, don't even have brains.Talking about consciousness is in any case a debate based on unproven assumptions.Where does consciousness come from?Who designed the programmer's program? The ultimate answer, says Holland, can only be that "no one is running this." Because if there is a programmer lurking behind the scenes, it's like "a ghost haunting the machine," then you have nothing to explain.You just pushed the mystery elsewhere.But fortunately, there is another option: feedback from the environment, which was Darwin's great insight.An agent can improve his own inner patterns without any supernatural guidance.It just keeps testing its models to see how accurate their predictions are about the real world.If it survives the practice, it adjusts these models of itself so that it does better next time.In biology, of course, the agent is a single organism, the feedback is provided by natural selection, and the continuous improvement of the model is called evolution.Cognitively, the process is basically the same: the agent is an independent mind, feedback comes from teachers and direct experience, and improvement is called learning.Indeed, this is exactly how Samuel's computer checkers player works.An adaptive agent, both biologically and cognitively, has to take advantage of what the world tells you. Of course, the next question is, how to do this?Holland discussed this basic concept at length with his colleagues in the BACH group.But in the end there was only one way to pin down the concept: a computer simulation of an adaptive agent had to be built, just as he had done with genetic algorithms fifteen years earlier. But unfortunately, he found that by 1977, the mainstream knowledge of artificial intelligence was not as helpful to him as it was in 1962.By 1977, the field of artificial intelligence research had undoubtedly made great progress.At Stanford University, for example, the artificial intelligence group is developing a series of highly productive programs known as expert systems.Expert systems can model expertise, such as a doctor, by running hundreds of rules: "If the patient has bacterial meningitis and is running a high fever, then maybe it's some kind of bacterial infection." This research has led to interest and attention of investors. But Holland was not interested in applied research.What he wanted was a basic theory of adaptive agents.From his point of view, the price of human progress in the field of artificial intelligence in the past two decades is that all important aspects have been ignored, from the study of learning to the study of feedback from the environment. Come on, feedback is the most fundamental problem.But apart from a few like Samuel, people in the field of artificial intelligence seem to think that learning can be put aside and taken care of.They thought they could wait until they had figured out their understanding of language, human problem-solving, or perfected programming of other abstract reasoning problems before studying the problem of learning.The designers of expert systems are even proud of it.They talk about "knowledge engineering," that is, after months of talking to relevant experts, developing hundreds of rules for a new expert system to answer: "What would you do in this situation? In that situation What are you going to do?" Questions like that. To be fair, even knowledge engineers have to admit that if programs can really learn their professional knowledge through teaching and experience like humans, if someone can figure out how to apply these software without being as complicated and troublesome as it is now , things will go much smoother.But for Holland, that's exactly the problem.Scrambling existing "learning models" into a piece of software won't solve any problems.Learning is the most fundamental problem of cognition, just as evolution is the most fundamental problem of biology.This means that the mechanisms for learning have to be put into the cognitive architectural blueprint at the outset, and not thrown in hastily at the end.Holland's ideal model is still the Hebbian neural network, the most important point of which is that every nerve impulse of thinking strengthens its neural connections, thus making thinking possible.Holland is convinced that thinking and learning are just two sides of the same coin in the brain.He hopes to capture this fundamental question in his work on adaptive agents. Still, Holland doesn't want to go back and redo the neural network simulations.Although it has been twenty-five years since IBM701, the computer is still not powerful enough to do a complete Hebbian computer simulation on the scale he wants to achieve.In the 1960s, neural network research did have a brief climax under the heading of "sense control".Perceptual control is a neural network dedicated to feature recognition in vision research.But perceptual control is an extremely simplified version of what Herb actually meant by the collection of cells. (Sense control is weak even at recognizing visual features, which is why it has fallen out of favor.) Holland is also not very impressed with the new generation of neural network systems.A new generation of neural network systems became popular in the late seventies and has received a lot of attention since.These systems are more advanced than visual control systems, Holland said, but they still don't support the study of cell assemblies.Indeed, most versions simply don't resonate.The signal waterfall through the network has only a single direction from front to back."These connectivist networks are powerful in terms of stimulus/feedback behavior and pattern recognition, but by and large ignore the need for internal feedback," he says. No. Except for a few cases, neural network researchers basically don't work on this aspect." As a result, Holland decided to design a hybrid simulated adaptive agent of his own, combining the strengths of neural networks and expert systems.In order to enhance computer efficiency, he started with the famous "if...then" rule of the expert system.But he adopted this rule from a neural network perspective. In fact, Holland said, there will be something like an "if...then" rule in any situation.In the late 1960s, long before expert systems were heard of, rule-based systems were introduced by Carnegie-Myron's Ellen Newell as computers for the common functions of cognition. and Herbert Simon introduced into computer programming.Newell and Simon treat each rule as a single knowledge package, or a single combination of techniques.Like "If the squeak is from a bird, then the squeak has wings", or "If the choice is between holding your opponent hostage or holding your opponent's wife, hold the opponent Mrs. Madame." Moreover, these rules point out that when programmers express knowledge in this way, the rules automatically acquire a certain wonderful flexibility of cognition.The rule of acting on conditions, "if this is the case, then do that," means that such systems do not operate within a fixed family, say, of some subroutines in FORTRAN or PASCAL.A particular rule is activated only when its conditions are met, so that its response is appropriate for its situation.Indeed, when a rule is activated, it is likely to cause a chain reaction of all rules: "If the situation is A, take action B", "If the situation is B, take C action", "If the situation is C, Just take action D", and so on.In general, a whole new program follows this series of chain reactions and will give perfect answers to the questions asked. Compared with the exciting game-like blind and rigid computer behavior, this It is really the mechanism needed by the intelligent system. In addition, rule-based systems have great implications for the neural distribution of the brain.For example, a rule is equivalent to one of the Hebbian collections of cells in a computer.He said: "In Heb's theory, a collection of cells is a simple statement: If events happen like this, then I will be fired at high speed for a while." The interaction of rules, with a rule activated The whole linkage to other rules that follows is like a natural consequence of a neurally densely connected brain. "Each of Heb's ensembles contains about a thousand to 10,000 neurons," Holland said. "Each neuron has another thousand to ten thousand synapses that connect to other neurons. So each cell assembly is connected to many other cell assemblies." Basically, activating a cell assembly is equivalent to Post a notice on an internal bulletin board and it will be seen by most, or all, other assemblies of cells in the brain. "Assembly 295834108 is now in action!" As soon as this notice appears, those assemblies that are appropriately associated with this collection will be activated and post their own announcements on the bulletin board, causing a constant repetition cycle. The internal organization of a Newell-Simon-style rules-based system is very close to this bulletin board metaphor, Holland said.The internal data structure of the system is equivalent to this bulletin board, which contains a series of digital notices.And then there's the huge set of rules that the computer encodes hundreds, even thousands of numbers into its own parts.While the system is running, each rule scans the bulletin board frequently, looking for a bulletin that matches its "if" criteria.Whenever one of the rules finds a notice that meets its own conditions, it will immediately post a data message to continue the "then" part. Holland said: "If you think of this system as a kind of office, there are memos on the bulletin board that must be dealt with today, and each rule is equivalent to a desk in the office, responsible for processing a certain kind of notice. At the beginning of the day, each desk collects the memos it is responsible for processing. At the end of the day, each desk posts the memos of the results on the bulletin board.” Of course, by the next day Start repeating the cycle again in the morning.In addition, some memos were posted by the probes to keep the system in touch with events taking place outside.Still other memos may be activated effectors, subsystems that enable the system to affect the outside world.The detectors and effectors are the computer equivalent of eyes and muscles, Holland said.So, in principle, a rule-based system can easily get feedback from its environment, which is one of its first requirements. So Holland used this similar bulletin board metaphor in his designs for adaptive agents, but at the same time he returned to defying conventional wisdom in detail. For example, in standard Newell-Simon terms, rules and memos posted on bulletin boards should be written in symbolic language like "bird" or "yellow" because we purposely Make it close to the concept in the human mind.For the vast majority of people in the field of artificial intelligence, the correctness of using such symbols to represent concepts in the human mind is beyond dispute. This has been the golden rule for decades. Newell and Simon are of this view The most eloquent representative.And, it does capture a great deal about how our minds actually think.Symbols in computers can be combined into cumbersome data structures to represent complex situations, just as concepts are associated with, and emerge from, patterns in a psychologist's mind.In turn, these data structures can be used by programs to compete with brain activities such as reasoning and problem solving, just as the types of simulations in our heads are reconstructed and changed during the process of thinking.Indeed, if you take Newell-Simon's argument literally, as many researchers do, you will see that this symbolic manipulation is thinking. But Holland just didn't buy it.“Symbol processing was a good start, and really a big step forward in understanding conscious thought processes,” he says. But the symbols themselves were too rigid, and there was too much left over.How can a single letter containing B-I-R-D (the English spelling of bird) data really capture all the subtle and erratic nuances?How could the letters have any real meaning to the program if they couldn't communicate with real birds from the outside world?Even leaving this question aside, where do these symbolic concepts come from in the first place?How did they evolve and develop?And how was it formed through feedback from the outside world? For Holland, this is precisely the lack of research interest in learning problems in the mainstream direction of artificial intelligence. "We're in the same dilemma we're in when we classify species before we understand how they evolved," Holland said. "You can learn a lot from systems like these, but in the end it's not very far." Still, he believes, the concept has to be understood from Herb's point of view: Emerging structures emerge from some more Developed from deep neural substrates that are constantly tuned in response to environmental feedback.Just as clouds form from the physical and chemical changes of water vapour, the concept is vague, erratic, and dynamic.They are constantly restructuring and changing shape. "The key to understanding complex adaptive systems is how did the layers arise?" he said. "If you ignore the laws of the next layer, you will never understand the problem at this layer." In order for his adaptive agents to grasp the concept of emergence, Holland decided that his rules and proclamations should not be written in symbolic means with special meaning.They will be just arbitrary sequences of binary ones and zeros.A bill might be a sequence like 10010100, similar to the chromosomes in his genetic algorithm.And a rule, in English, might be like: "If there's a notice on the bulletin board that reads 1###0#00, where # means 'don't care,' put the notice 01110101 on it." This notation was so unconventional that Holland had to give his rules a new name, "classifiers," because their approach was to classify different notices according to their particular type.He thinks this abstract notation is crucial.Because he sees too many AI researchers fooling themselves, pretending that their symbol-based programs "know".In his classifier system, the meaning of the announcement must come from the way it causes one classifier rule to activate another classifier rule, or it has meaning because some part of it is directly detected by a real-world sensor written by.Concepts and mental models then emerge as a self-supporting swarm of classifiers that should be able to self-organize and reorganize like an autocatalytic group. At the same time, Holland draws exceptions to the conventional notion of central control in rule-based systems.Common sense rule-based systems are too flexible, so some form of central control has to be devised to avoid anarchy.Because there are hundreds of rules vying to see the bulletin boards filled with notices, there will always be several rules popping up and fighting each other over who should post the next notice.Assume that it is impossible for all the rules to post the next notice, since their notices may be completely incoherent ("Take hostage" or "Seize his wife"), or their notices may cause entirely different cascades of rules, such that It will cause the whole system to behave completely differently.So, to prevent computer schizophrenia, most systems implement cumbersome "dispute resolution" strategies to ensure that only one rule can take action at a time. But Holland argues that this top-down approach to dispute resolution is precisely what's wrong.Is the world so simple and so predictable that you can always know in advance what the best rules are?almost impossible.And if the system is told how to act in advance, it would be a lie to call it artificial intelligence: such intelligence is not in the program, but in the programmer's head.No, what Holland wants is to let control come from learning.He wants control to emerge from the bottom, as emerges from the neural substrate of the brain.To hell with continuity.If the two classifier rules cannot agree on each other, let them compete for a result based on their own performance. This result is a proven contribution to the task, rather than a software designer pre-programmed Good program selection. “Contrary to mainstream AI research, I think competition is more essential than coherence.” Coherence is an illusion, because in a complex world, coherence of experience is not guaranteed.But for an agent playing a game with his environment, competition is constant. "Furthermore, we have not distilled out the main qualities of competition, other than studies of competition in economics and biology." We are only just beginning to appreciate the richness of competition.Think of the amazing dynamics that competition can generate to cooperate, where certain actors spontaneously form alliances for mutual support and form symbolic relationships with each other.This happens at every level of all complex adaptive systems, from biological to economic to political. "Competition and cooperation may seem like opposites, but at some deep level they are two sides of the same coin." In order to realize the mechanism of competition, Holland decided to turn the posting into some kind of auction.His basic idea is to think of classifiers not as computer instructions, but as assumptions and guesses about what is best to post in a given situation.The rationale and strength of each assumption is measured by the value of each hypothesis, so that there is a basis for bidding.In Holland's concept of posting, each cycle begins as before, with all the classifiers scanning the bulletin boards for posts that are relevant to them.As usual, they will stand up and prepare to post their own notices as soon as they find a classifier related to them.But instead of posting their own notice immediately in the past, each classifier will first bid according to its ability.A classifier that is confident in the experience of "the sun will rise in the east tomorrow" might bid a thousand, while a classifier that is convinced that "Elvis is alive and at 6 Walla Walla at night" might bid $1,000. one.然后这个系统就会收集所有的出价，用抽彩给奖法选择一组赢家，叫价最高的最有可能赢。中选的分类器就会张贴它们的布告，就这样循环往复。 复杂吗？荷兰德不予否认。而且，这种拍卖就是以任意可信值取代任意争议解决战略。但现在让我们假设这个系统能够从这些可信值中吸取经验，那么这种拍卖就会排除中央仲裁人，从而让荷兰德获得他正想要的东西。并不是每一个分类器都能够赢：布告栏非常大，但却并非无限大。也并不是跑得最快的就一定能赢。如果时来运转的话，即使“猫王还活着”也能得到张贴自己的布告的机会。但一般来说，总是那些最强健的和可信值最高的假设获得系统表现的控制权，而那些离谱的假设时不时出现只增加了这个系统的一点儿自发性。而且如果有一些假设相互矛盾，那也不应该成为危机，而应该是一个机会，一个让系统决定谁的可信度更大，从而吸取经验的机会。 所以，我们又返回到学习这个问题上来了：分类器怎么来证明自己的价值，又怎么为自己获取可信值呢？ 对荷兰德来说，最显在的答案就是采用一种赫伯式的强化作用。每当一个作用者做对了什么事，从环境中得到了一个正反馈，它就应该强化那些与此相关的分类器。而每当它做错了什么事，它就同样应该削弱相关的分类器。无论采取强化的方法，还是采取削弱的方法，它同时应该不去理会那些不相干的分类器。 当然，关键是要弄明白这些分类器所起的作用。作用者不能奖赏那些在颁奖的时候正巧表现活跃的分类器。那就会像把得分的一切功劳都归于那个凑巧带球冲过底线的队员，而对操纵全局、把球传给他的四分卫，对拦截了对方进攻、为他开路的前锋，或任何替他传球的队员的功劳一笔勾销了。这也像把赢得一盘国际象棋的全部功劳都归于将住了对手的国王的最后一步棋，而无视为获得全局胜利而布局中的许多关键的棋步。但还有其它替代办法吗？如果作用者为了奖赏正确的分类器而必须预期回报，在没有被编入程序的情况下立该怎么做呢？在事先一无所知的情况下，作用者如何得知这些布局的价值呢？ 这确实是一个问题。不幸的是，赫伯式的强化作用是一个过于广泛的一般性概念，无法提供解答。荷兰德感到非常困惑，一直到有一天他偶然回想起他在麻省理工学院上的基本经济学课程，那是著名的经济学教科书撰写人保尔·塞缪尔森上的课，他才意识到他几乎已经解决了这个问题。他的布告栏前的拍卖已经为他在系统中建立了某种市场机制，通过允许分类器量力叫价的办法，他已经创造出了通货。所以，为什么不采取下一步行动？为什么不创造一个完整的自由市场经济，让强化能够在利益驱动下发生作用呢？ 确实，为什么不呢？当他终于看到了这一点，就发现这与经济非常相似。荷兰德认识到，如果把张贴在布告栏上的布告当作是上市叫卖的货物和市场上提供的服务，那么就能把分类器想成是生产这些产品和提供这些服务的公司和厂家。当一个分类器看到有一个布告满足了它的“如果条件”，它就会叫一个价，那么就可以把它想成是一个正在求购生产所需供应的厂家。为使这一相似性更加完善，他要做的是，必须要使每一个分类器对自己消耗的供应付出报酬。他决定，当一个分类器赢得了张贴自己的布告的权力，它就得将自己的一部分力量转给供应商，也就是那些触发其张贴布告的分类器。在这个过程中，这些分类器就会被削弱。但在下一轮拍卖中，一旦它的布告上市，它会有重新聚集力量的机会，甚至能够获利。 但这些财富究竟是从何而来的呢？当然是从最终消费者而来的：环境就是系统的所有报偿之源。荷兰德认识到，除此之外，对凑巧在颁奖的时候活跃异常的分类器给予奖赏是完全正确的。既然每一个分类器都对供应有所付出，那么市场就会保证其奖赏普及到所有中选的分类器，从而产生他所寻求的某种自动报偿和惩罚机制。他说：“如果你生产出对大家都合适的产品，那么你就会获利。如果不是这样的话，那就没人会买你的东西，你就会破产。”所有能够产生有效行动的分类器都会被强化，任何参与布局的分类器都不会被忽略。随着时间的日积月累，随着整个系统不断汲取经验和从环境中获得反馈，每一个分类器的强度就会与自己对作用者的真正价值相符。 荷兰德将适应性作用者的这部分称为“水桶队列”算法，因为其方法是将奖赏从一个分类者传到前一个分类者。这有如希伯的强化神经突触的大脑理论的直接翻版。或者，从这个意义上来说，与在计算机上调训模拟的神经网络也如出一辙。当想到这些时，荷兰德知道他快要触及到问题的实质了。以利益为驱动力的经济强化是一个极为强大的组织力量，就如亚当·斯密的那只看不见的手在现实经济中具有强大的力量一样。荷兰德认识到，从原则上说，你可以用一组完全随意的分类器来启动系统，这样，作用者这个软件就会像新生婴儿一样手舞足蹈地乱蹬乱踹。然后，随着环境不断强化某些行为，随着水桶队列发生作用，你可以看到分类器将自己组织为前后连贯的序列，从而产生预期的行为表现。一句话，学习从头开始就被设入于系统之中了。 这么说，荷兰德几乎就摸到门了，但还不尽然。荷兰德把水桶队列算法建立在基本的基于法则的系统之上，并赋予了他的适应性作用者某种形式的学习功能。但适应性作用者还缺少另一种学习的形式，开采式学习与探险式学习之间是有区别的，水桶队列算法能够强化作用者已有的分类器，可以打磨已有的技能，但它却无法创新。仅仅只依靠水桶队列算法，会使系统趋于最大化的平庸状态，因为这个算法无法使系统在无限广阔的可能性空间搜索到新的分类器。 荷兰德认为，搜索于可能性空间正是基因算法可以承担的工作。事实上，当你想到这一点时你就会看到，达尔文的比喻和亚当·斯密的比喻恰好可以相辅相成：企业能够随时间进化，为什么分类器不能够呢？ 荷兰德当然不会为这一洞见而大惊小怪：基因算法一直存储在他脑子里。他刚开始对分类器做二进制的表述时就想到了基因算法。分类器用英文来陈述就像：“如果有两个布告，其模型分别是1＃＃＃0＃00和0＃00＃＃＃＃则张贴布告01110101。”但在计算机里，各部分信息会被串在一起，被写成一连串的信息： “1＃＃＃0＃000＃00＃＃＃＃01110101”。对基因算法而言，这就像是数字染色体。所以可以完全用同一种方式来执行这个算法。在大多数情况下，分类器会像以往一样在市场上欣然买进或卖出。但系统会经常性地选择最强的一对分类器来繁衍后代。这些中选的分类器会通过性交换来产生一对后代，从而重组它们的数字化建设砖块。而新生代会取代一对力弱的分类器。然后，新生代将有机会来证明自己的价值，通过水桶队列算法使自己越变越强壮。 结果就是，这群规则会随时间而改变和进化，在可能性空间中不断发现新的领域。由此你就会达到目的：将基因算法当作第三层，置于水桶队列算法和基本的基于规则的系统之上，荷兰德终于构筑成了一个不仅能够吸取经验，而且具有自发性和创造性的适应性作用者。 他现在所要做的就是，将这个构想变为一个可以运作的软件程序。 荷兰德从1977年左右开始为第一个分类器系统编码。奇怪的是，这项工作并不像他期望的那样直截了当。“我真以为只消几个月时间我就可以编出能够运作、对我有用的程序来。但实际上，我用了大半年的时间才做到令我自己满意的地步。”他说。 另一方面，这也要怪他自己让自己做难。他以真正的荷兰德的风格来编写第一个分类器系统：完全依靠自己，而且是在家里，用的是十三年前他用于旋风计划的十六进位码和他家的一台康莫多（Commodore）计算机。 巴奇小组的成员们至今在说到这一段故事时还带着诧异的神情。当时满校园都是计算机：VAX机、大型计算机、甚至高功能的绘图工作站。为什么要用康莫多机？为什么要用十六进位码？几乎没人还在用十六进位码了。如果你真是个死心塌地的计算机高手，想方设法要从一台计算机的程序中榨出最后一点利用价值的话，你也可以用所谓组合语言来写，那起码能够用像MOV、JMZ和SUB这样的帮助记忆的符号来取代数字。或者，你也可以用PASCAL、C、FORTRAN或LISP这样的高级语言来编写程序。这些语言是人类比较容易理解的。尤其是科恩，仍然记得为此与荷兰德做过长时间的激烈争论。如果用这些夹杂字母的数字将程序写得杂乱无章，谁会相信它能运作呢？就算有人相信你，但如果你的分类器系统是在家用计算机上编写成的，谁又会用它呢？ 最终荷兰德只好做出让步。不过到他同意将分类器系统交给一个研究生，里克·里奥罗（Rick Riolo）时，早已是八十年代初了。里克将这个系统改编成一个一般性功能的、能够在所有类型的计算机上运行的软件系统。荷兰德承认说：“只不过这不是出于我的本能。我总是喜欢将实验做到能让我看到它真的能够运行的地步，然后就失去了兴趣，又回到了理论。” 正因为如此，所以他仍然坚持认为当时的康莫多计算机对他而言意义甚大。大学的计算机上是共享的，这令人头痛，他解释说：“我喜欢直接在计算机大忙小乱地编写程序，但如果用大学的计算机，就不可能容许我一口气上机八个小时不下来。”荷兰德把个人电脑看作是上帝的恩典。“我发觉我可以在我的个人电脑上编写程序，可以独个拥有于家中，不用再依赖任何人。” 另外，为旋风和IBM701编写程序的经验使荷兰德一点儿也不觉得这些桌面小电脑过于原始。事实上，当他买康莫多电脑时，他觉得已经前进了一大步。实际上他1977年就买了被称为“微心智”（Micromind）的电脑，当时这台电脑看上去像是崭新的苹果二型机的强劲对手。“那是一台很好的小型计算机。”他回忆说。确实，尽管那只不过是一个黑匣子里的一堆电路板，可以与电报打字机连接后做信息输入和输出，而且没有屏幕，但它有8000字节和8位内存。价值三千美元。 说到十六进位码，那是因为微心智计算机当时没有其他语言可供使用，而荷兰德又不愿意等待。“我习惯了用组合程序，我用十六进制码就像用组合程序一样方便，所以用十六进制码来写程序对我来说并不困难。” 讲完这段故事后，荷兰德说，微心智计算机公司这么快就倒闭了，真是太令人遗憾了。他一直到感到八千内存实在不够用时才开始转用康莫多电脑。他说，当时康莫多电脑是一个最理想的选择。它采用了与微心智同样的微处理器芯片。这意味着，几乎不用做任何改变就能够让它运行十六进制码。康莫多的内存要大得多，带屏幕显示，而最大的好处是，“康莫多能让我玩游戏。”他说。 尽管荷兰德的同事对此非常恼火，但他的第一个分类器系统却运转得非常好，这足以能够使他确信，这个系统确实实现了他的意图，而且确实为完整的认识理论播下了种子。这个系统的早期版本是他与密西根大学心理学教授裘迪·瑞特曼（Judy Reitman）共同研制的，发表于1978年。在对这个版本系统的测试中，其作用者学会了如何用基因算法运行一个模拟的迷宫，运行速度要比没有用基因算法快十倍。这次测试同时也证实了，分类器系统能够显示心理学家所称的“转换”：它能够把在前一个迷宫中学到的规则运用到后一个迷宫的运行中去。 这些早期研究成绩斐然，即使荷兰德并不大肆渲染，其名声也已使“分类器系统”这个词开始流行了起来。比如1980年，匹兹堡大学的史蒂芬·史密斯（Stephen Smith）开发了一个能够玩扑克的分类器系统，并用它来和一个也有学习功能的老一点的玩扑克牌的软件对抗。这场对抗甚至不成其为比赛，分类器系统轻而易举地就赢了。1982年，泊拉罗德公司的斯图尔特·威尔逊（Stewart Wilson）用分类器系统来协调电视摄影机和机械手臂的动作。他的应用表明，水桶队列和基因算法能够导致分类器规则的自发组织，从而自我分类成一个个小组，起到控制子规则的作用，产生我们所需要的特殊而协调的动作。1982年，荷兰德的学生拉森·勃克（Lashon Booker）在他的博士论文中将分类者系统运用在一个模拟的环境，用它来寻找“食物”，避免“食物中毒”。这个系统很快就将自己的规则组织成这个环境的内化模型，就像一幅心智地图。 但对荷兰德来说，最感欣慰的是1983年戴维·高德勃格（David Goldberg）的研究证明。高德勃格是一个攻读博士学位的土木工程师，在此几年前就选修荷兰德的适应性系统课程，并一直对此深信不疑。高德勃格说服了荷兰德成为他的博士论文答辩委员会的主持人之一。他的论文证明了，怎样把基因算法和分类器系统运用于对一个模拟的煤气管网线的控制。当时，这是分类器系统对付过的最复杂的问题。任何一个煤气管道系统的目的都是以尽可能小的成本来满足终端用户的需要。但每一条煤气管道都有几十个或几百个压缩机，将煤气从几千英里的大管径管道抽出来。用户的煤气用量每一小时、每个季度都会有变化，而压缩机和管道常会有渗漏，限制了整个系统在适当压力下的供气能力。安全控制要求煤气的气压和运速保持在适当的程度，但任何一个因素都会影响到其它因素，即使想使一个简单的煤气管道发挥最大效益，都复杂到根本无法用数学来分析。管道操作者是通过长时间的“学徒”才学会用本能和感觉来控制煤气管道系统的这门技术的，就像我们学开车一样。 事实上，煤气管道的问题之复杂，就连荷兰德都发愁，担心高德勃格的研究较之其它分类器系统运用小组，也许更可能失败。但其实他根本不必担心。高德勃格的系统非常圆满地学会了控制这个模拟的煤气管道系统：这个系统从一组完全随意的分类器开始，在经过一千天的模拟试验之后，达到了对控制煤气管道的专家水平。而且，这个系统掌握操作煤气管道的规则简单得不可思议。它的布告仅为十六个二位数那么长，它的布告栏上每次只有五条布告，总共只有六十条分类器规则。事实上，高德勃格在他家的苹果二型机上只用了64千字节的内存就运作了整个分类器系统和煤气管道的模拟程序。荷兰德笑着说：“高德勃格是我最紧密的追随者。” 煤气管道的模拟不仅使高德勃格在1983年获得了博士学位，而且使他获得了1985年度的总统青年研究者奖。荷兰德自己也将高德勃格的研究看作是分类器系统研究的一个里程碑。“这非常有说服力，”他说。“它真正解决了一个实际问题，或者说，起码是解决了一个实际问题的模拟。”而且，不无讽刺也不无欣慰的是，这个分类器系统最“实际”的一例，反过来对基本的认知理论也最具说服力。 荷兰德说，这一点在高德勃格的系统如何学会控制渗漏的方法中表现得最为明显。这个系统从一组随意的分类器开始，首先掌握一系列对正常的煤气管道的运行较为广泛适用的规则。比如在一次传送煤气的操作中，出现了一条可以被解释为“一直传送'没有渗漏'的信息”。很显然，这是一条过于一般的规则，只适用于管道运行正常的情况。但在高德勃路开始在各种模拟的压缩机上打出模拟的洞来的时候，这个系统很快就发现了这个问题，其操作立刻就失灵了。但通过基因算法和水桶队列，这个系统最终从自己的错误中反省了过来，开始产生了一些比较特殊的规则，比如“如果输入气压很低，输出气压也很低，气压转换率是消极的，则传送'渗漏'信息”。而且，只要这条规则一经实行，就会产生比第一条规则高得多的叫价，把第一条规则从布告栏上取代下来。so and so.大体上说，第一条规则在不出现非常规行为的正常情况下会发生作用，而一旦发生意外情况，第二条规则和其他规则就会取代第一条规则，对意外的行为做出校正反应。 当高德勃格告诉荷兰德这些时，荷兰德激动万分。在心理学上，这种知识的组织被称为缺席的等级制度（defulthierarchy），当时这正是荷兰德久埋脑海的研究课题。自1980年起，他一直在与三位密西根大学的同事，心理学家凯瑟·赫力约电（Keith Holyoak）、里查德·尼斯伯特（Richard Nisbett）和哲学家保尔·查加德（Paul Thagard）密切合作，致力于创立一个关于学习、推理和知识发掘的认知理论。正如他们在1986年出版的《归纳法》一书中所说的那样，他们四个人都相信，这个理论必须建立在三项基本原则上，而这三项基本原则也正是荷兰德的分类器系统的原则：即，知识能够以类似规则的思维结构来表达；这些规则始终处于竞争之中，经验使得有用的规则越变越强，无用的规则越变越弱；具有说服力的新规则产生于旧规则的组合之中。这个有大量的观察和实验结果支持的观点表明，这些原则可以解释各种恍然大悟的洞见，包括从牛顿对苹果落地的顿悟，到日常生活中对相似性的发现。 他们特别指出，这三项原则应能够产生缺席的等级制度的自发涌现。确实，这正是人类所有知识的基本组织结构。一组规则形成缺席的等级制度，从根本上来说与荷兰德所称的内在模型是同一个意思。我们用较弱的一般性规则和较强的特例来预测事物该如何分类：“如果它是流线型的，有鳍，生活在水中，那它就是鱼。”但“如果它同时还有毛发，呼吸空气，而且很大，那它就是鲸。”我们用同样的结构来预测如何做事：“'i'总是在'e'之前，除非其后有'c'。”但“如果是neighor、weigh、或weird这些字，则'e'总是在' i'之前。”我们还用同样的结构来预测因果关系：“如果你朝一条狗吹口哨，它就会向你跑来”，但“如果它向你嗥叫，并把颈背部的毛发竖起来，那它也许不会向你跑来。” 荷兰德说，这个理论说明，无论这些原则是作为分类器系统来执行，还是以别的形式来实行，缺席的等级制度都应该会涌现出来。（事实上，《归纳法》一书中引用的许多计算机模拟实验都是用PI来做的。PI是查加德与赫力约克设计的更常规的基于规则的软件程序。）不管怎么说，看到等级制度能够真的从高德勃格的煤气管道模拟中涌现而出，真是非常令人激动。分类器系统总是从零起步，它最初的规则完全是在计算机模拟的太初混沌中随意设置的，然而就在这混沌之中，美妙的结构涌现了出来，令人惊喜，让人讶异。 “我们感到欢欣鼓舞，这是能够真正被称为涌现模型的首例。”荷兰德说。